When Gousto hands you lemons…we create Recipe Performance targets.

Neal Patel
Nov 12 · 4 min read

Catchy title right? Yeah I didn’t think so either, but I spent more time than I care to admit thinking of it so it’s only right I use it! Here at Gousto we’re in a fortunate position where we can get creative with our data to come up with new ways of understanding customer feedback. The ways we utilise data are endless so that’s probably one for a separate blog, but the metric that I want to dive into today is the Recipe Performance Scores (RPS).

For context, after customers receive a recipe we ask them to rate it between 1 and 5 stars which helps determine the popularity of a recipe. These recipe ratings are then used to segment customers into “Promoters” (five star ratings) and “Detractors” (three star ratings or less) which feeds into a simple calculation:

((Detractors — Promoters)/Total Number of Respondents)*100

The result gives us a value between 100% and -100% (which admittedly is very rare) and thus the Recipe Performance Scores are born.

Historically, we’ve had one standard RPS figure of 35% that we should be aiming for across all recipes, which is a good start but there was never any analytical backing behind this figure. For example “Miso Tofu” generally has an average RPS of 33% whereas a “Bacon Cheeseburger” has an average RPS of 68%…so it doesn’t make sense for both these recipes to be measured against the same static 35% target.

Visualisation demonstrating how we use average RPS scores per ingredient to create our RPS targets.

I bet you’re wondering where the “lemons” come into it? Well that’s to do with how we now set our Recipe Performance targets (RPS targets). Instead of having a static number, we decided to split out all of our recipes by their most polarising ingredients (e.g. lemons) using the expertise of our talented recipe developers who know which ingredients tend to perform well & poorly. By isolating these ingredients, we can then get an average for all recipes including each of these individual ingredients over the last six months. Since the ingredient composition for each recipe is different, we end up with unique RPS targets per recipe to benchmark against. As an example, recipes with beef mince are generally 11% higher than the base but if the recipe also had sweet potatoes we would also have to subtract 4%. This would be calculated using all polarising ingredients for a recipe which would give us a unique RPS Target for each recipe.

Visualisation showing how RPS targets translate to a Relative RPS score.

To supplement the Recipe Performance targets, we also created a Relative RPS metric which gauges how well a recipe has done versus its target in the form of an index. As an example, a relative RPS score of 1.00 indicates that the recipe performed as expected whilst a figure above 1.00 indicates it performed better than expected and so forth. This adds another lens to the way we are able to make decisions on what our customers thought about our recipes & whether they’re ready to go on the menu again or need to be reimagined/banned.

RPS targets & Relative RPS have already had a massive impact on how they’re helping our recipe developers understand our customers more. Day to day they’ve helped inform recipe developers on; whether we can put recipes back on a menu, whether certain ingredients are underperforming and if menus are performing as well as we’d expect them too. It’s also been pivotal in the menu planning process & in post-collection analysis for campaigns, such as Seoul Train & Wild Buns, where it makes up one of the six key themes that we use to measure campaign performance.

For those that have made it this far (thank you), you’ll be glad to know that this isn’t the end of RPS but it’s actually just the beginning. Such is the Gousto culture, we’re already looking at working with data science to see if we can tie in some of their handy-work to take the RPS targets to that next level — so watch this space!

Gousto Engineering & Data Blog